Add LocalLLM/vLLM provider support

From the PR description

Overview

Adds support for OpenAI-compatible local LLM endpoints (vLLM) as a third provider alongside Claude and Gemini. This enables users to connect Mike to self-hosted LLM instances while maintaining compatibility with commercial APIs.

Key Changes

Backend:

• New openai.ts provider adapter for OpenAI-compatible endpoints
• Configurable model names via environment variables (VLLM_MAIN_MODEL, VLLM_LIGHT_MODEL)
• Server-side configuration (no user API key management required)
• Added fallback for missing title parameter in generate_docx tool

Frontend:

• Added "LocalLLM" model group to model selector
• Default models: LocalLLM Main and LocalLLM Lite
• Updated model availability checks to treat LocalLLM as always available
• Updated account settings page to reflect LocalLLM configuration

Configuration:

• New environment variables:
  • VLLM_BASE_URL - Endpoint URL for OpenAI-compatible API
  • VLLM_API_KEY - Optional API key for authentication
  • VLLM_MAIN_MODEL - Model name for main chat tasks
  • VLLM_LIGHT_MODEL - Model name for lightweight tasks (title generation)

Testing

• Tested with vLLM endpoint at https://bredaai.com/api/v1
• Verified document generation workflow
• Confirmed model selection and routing works correctly

Notes

• LibreOffice is required for document conversion (DOCX ↔ PDF)
• LocalLLM models are server-configured and available to all users
• Maintains backward compatibility with existing Claude/Gemini providers